ai-agents-config-catalog

name	ai-agents-config-catalog
description	Catalog of every configuration axis in this repo, env vars, commit markers, frontmatter keys, QA skip verdicts, and escape hatches, each with its enforcement point and abuse story, plus the checklist for adding a new flag safely. Use when you say `what does a skip flag do`, `list escape hatches`, `can I skip this gate`, `add a config flag`. Do NOT use for hook runtime behavior (use `agent-harness-reference`) or change gating policy (use `ai-agents-change-control`).
version	1.0.0
license	MIT

AI Agents Config Catalog

Every flag, marker, and skip semantic in this repo, verified against code as of 2026-07-03. Each escape hatch exists because a gate sometimes misfires; each one also has an abuse story or a guard. Before you set any of these, read its row. The house rule (learned in session 1187, see the Removed Flags section): escape hatches get teeth or get abused.

Related skills: ai-agents-change-control owns when a bypass is allowed; agent-harness-reference and ai-agents-architecture-contract own what the hooks themselves do; ai-agents-debugging-playbook owns triaging a gate that fired on you.

Triggers

• what does the skip-drift-check marker do
• list escape hatches
• can I skip this gate
• add a config flag
• is this skip marker allowed

Process

1. Identify the axis type: env var, commit marker, text directive, frontmatter key, file, or verdict string.
2. Find its row in the tables below. Read the effect AND the guard/abuse column before using it.
3. If you are about to use an escape hatch, confirm the legitimate trigger condition holds (for example, a workflow actually unrunnable under act). Bypassing because a gate is slow or annoying is the session 1187 failure mode.
4. Re-verify the flag in its defining validator with the one-liner in Provenance. Documentation can drift after validator changes.
5. Adding a new flag? Follow the checklist in "How to Add a New Flag" and add a row plus a re-verify one-liner to this catalog.

Environment Variables: LSP Gate (retired)

The LSP-first runtime gate and its three environment escapes (SKIP_LSP_GATE, LSP_GATE_MODE, LSP_DOWN) were retired in #3216 when ADR-062 was amended to keep LSP-first navigation as static steering only. No environment variable governs it now; the guidance lives in .claude/rules/lsp-first.md. Nothing to set, nothing to escape.

Git Hook Configuration

Lefthook is the sole Git hook manager. lefthook.yml declares the events, filters, named jobs, and validator commands. Do not infer a bypass from a deleted custom payload. Inspect the named validator's current interface before using or documenting an environment variable.

Removed and Stale Flags (do not use, do not reintroduce)

Name	Status as of 2026-07-03	Story
SKIP_PREPUSH	REMOVED	Historical: abused 3x within hours of creation (session 1187, retro .agents/retrospective/2026-02-08-session-1187-skip-prepush-abuse.md)
SKIP_TESTS	REMOVED	Historical: stale contributor guidance was removed during the Lefthook migration

Lesson encoded: a global bypass with no teeth (no telemetry, no approval step) will be reached for under pressure. New escape hatches must be narrow, announced in output, and observable.

Environment Variables: Tests and Tooling

Name	Type	Effect	Status	Guard / abuse story	Where defined
GIT_CONFIG_COUNT + GIT_CONFIG_KEY_0/GIT_CONFIG_VALUE_0	env vars (set by conftest)	Test session injects commit.gpgsign=false with command-line precedence so test repos never invoke the user's signing setup	Production test infra	Only sets index 0 when GIT_CONFIG_COUNT is unset, so an outer process's config is not clobbered	tests/conftest.py:35-38
PEP 668 / uv	environment reality	Bare pip fails (externally managed env). Everything goes through uv sync --frozen --extra dev; skill scripts need uv run python, not python3 (PyYAML lives in the venv)	Production	ModuleNotFoundError: No module named 'yaml' means you used the wrong interpreter	pyproject.toml, .python-version (3.14.6)
pytest markers unit, integration, safe_push_transport, security, smoke, windows_path	pytest -m selectors	Filter test classes; smoke = real-CLI tests needing auth/credits, nightly only, and the smoke gate asserts they were NOT skipped (issue #2231 item 4); safe_push_transport = touches a non-local transport, excluded from pre-push	Production	Marking a test smoke to dodge CI is detected by the not-skipped assertion	pyproject.toml [tool.pytest.ini_options].markers

.env Keys (from .env.example)

Copy .env.example to .env. Keys as of 2026-07-03: ANTHROPIC_API_KEY (MCP servers, skill-learning hook LLM fallback, exact token counts), PERPLEXITY_API_KEY, TAVILY_API_KEY, CONTEXT7_API_KEY, YDC_API_KEY, and optional COMPRESS_TOKENIZER (tiktoken/anthropic/heuristic, default tiktoken). Everything degrades gracefully when absent per ADR-007; missing keys disable the corresponding MCP or fallback path, they do not break the repo. Note: COMPRESS_TOKENIZER appears only in .env.example; a code consumer was not located in this audit, treat as possibly vestigial.

Commit Markers

Name	Type	Effect	Status	Guard / abuse story	Where defined
[skip-drift-check]	commit message marker	Skips the agent drift detection CI gate for the whole PR (marker in ANY commit subject/body counts)	Production escape hatch with obligations	The bypass job posts a checklist that a human must satisfy: reason documented in PR description, templates/README.md updated with the intentional difference, explicit code-owner approval. Marker alone is NOT approval	.github/workflows/agent-drift-detection.yml:65-69 (detection), :297-320 (obligations); CONTRIBUTING.md:499

Text Directives (orphan-ref-validator)

Name	Type	Effect	Status	Guard / abuse story	Where defined
orphan-ref-ignore HTML comment	line directive	Mutes orphan-reference findings for that one line	Production	Rare in tree, easy to grep; compensating control is human review of any commit adding one (DESIGN-009)	.claude/skills/orphan-ref-validator/scripts/patterns.py:95
orphan-ref-ignore-file HTML comment	file directive	Mutes the whole file, but ONLY if the directive appears within the first 50 lines	Production	The 50-line rule prevents burying the mute at the bottom of a long doc	patterns.py:96; 50-line window at scan.py:157

Write the directives as HTML comments (<!-- ... -->); shown bare here so this catalog does not mute itself.

Frontmatter Keys (skills)

Name	Type	Effect	Status	Guard / abuse story	Where defined
size-exception: true	SKILL.md frontmatter	Exempts the file from the 500-line block in the size validator	Escape hatch, must be justified	taste-lints only honors it in the first 20 lines of the file; validator prints "size-exception declared" so reviewers see it	scripts/validation/skill_size.py:107,251; .claude/skills/taste-lints/scripts/taste_lints.py:449
allowed-tools, argument-hint	frontmatter keys	Harness metadata; recognized (not flagged as prose refs) by the orphan-ref scanner	Production	Whitelist lives in one place	.claude/skills/orphan-ref-validator/scripts/filters.py:31

QA Skip Verdicts (ADR-034)

Session-end QA can be skipped only with one of these exact verdict strings in the session log, and only when the staged files qualify.

Verdict	When legitimate	Enforcement
SKIPPED: investigation-only	Every staged file matches the allowlist: .agents/sessions/, .agents/analysis/, .agents/retrospective/, .serena/memories/, .agents/security/, .agents/memory/ (incl. episodes/), .agents/architecture/REVIEW-*, .agents/critique/	Single source of truth scripts/modules/investigation_allowlist.py; pre-check via the session skill (Test-InvestigationEligibility); CI backstop .github/scripts/validate_investigation_claims.py (advisory)
SKIPPED: docs-only	All changes are markdown and strictly editorial: spelling, grammar, formatting; no code, config, tests, workflows, or code-block changes	SESSION-PROTOCOL.md:754; CONTRIBUTING.md:697

Mixed sessions do not qualify; split the commit. Claiming investigation-only with a code file staged is exactly what the CI backstop exists to catch.

Plugin Manifest Version Prohibition

Not a flag, but the config obligation most often tripped over. No packaged plugin manifest may carry a version field, and neither may a marketplace entry. Claude Code resolves freshness from the first version it finds, so a committed version pins consumers to a hand-bumped string instead of the git commit SHA (ADR-092, which supersedes ADR-079; issue #4080).

File	Must not carry
.claude/.claude-plugin/plugin.json	version
src/claude/.claude-plugin/plugin.json	version
src/copilot-cli/.claude-plugin/plugin.json	version
.claude-plugin/marketplace.json (per entry)	version
.github/plugin/marketplace.json (per entry)	version

Enforced locally by the pre-pr-validation job in lefthook.yml, which runs scripts/validation/pre_pr.py; that runner invokes build/scripts/validate_plugin_version_bump.py. CI also enforces it through .github/workflows/validate-plugin-version-bump.yml, which fires on a plugin source dir or a marketplace file. A content change needs no manifest edit at all.

Hook Registration Surfaces

Two independent registration sources serve different consumers. Do not force parity between them:

Surface	Consumer	Shape re-verified 2026-07-22
.claude/settings.json	Claude Code direct in this repository	5 events, 7 groups
.claude/hooks/hooks.json	Vendored plugin source for both harness packages	2 events, 2 groups

The Copilot generator reads .claude/hooks/hooks.json, not local settings. A one-file registration is valid only when its consumer scope is deliberate. Check both sources in any hook PR and document repository-only versus vendored intent. No PermissionRequest policy hook is registered. Test runners execute repository-controlled code, so command-name matching is not a safe approval boundary.

How to Add a New Flag

• Define it in exactly one module (single source of truth). Name it for its scope: SKIP_ACTIONLINT not SKIP_CHECKS.
• Document semantics including the failure mode: what happens when unset, when the guarded thing is broken, and whether the consumer fails open or closed. There is no neutral default for a missing signal (FM-10, PR #1965).
• Make every activation observable: print a WARN/SKIP line or emit EVENT= telemetry (guard-maturity consumes these). A silent bypass is the session 1187 pattern.
• Scope it to one check. Global bypasses (SKIP_PREPUSH) are banned by precedent.
• Add tests: positive (flag honored), negative (flag absent = enforced), edge (bad value), per TESTING-RIGOR.
• If the flag lives under .claude/, regenerate the src/copilot-cli/ mirror. Do not touch plugin.json; it carries no version.
• Add a row to this catalog and a re-verify one-liner to Provenance below.
• Route the change through ai-agents-change-control classification (a new escape hatch is never docs-only).

Anti-Patterns

Anti-pattern	Why it fails	Do instead
Reintroducing a global bypass	Session 1187: abused 3x in hours; user verdict "You can't be trusted"	Narrow, announced, per-check escapes
[skip-drift-check] without the checklist	Marker skips the CI job but the bypass job posts unmet obligations; reviewers will bounce it	Document reason, update templates/README.md, get code-owner approval
Documenting a flag only in CONTRIBUTING.md	Docs drift previously left removed flags in active guidance	The defining script is the source of truth; docs quote it (FM-9)
Editing .claude/lib/hook_utilities/ to change flag behavior	That tree is a generated mirror; next sync_plugin_lib.py run reverts you	Edit scripts/hook_utilities/, run the sync
Claiming SKIPPED: investigation-only with code staged	CI backstop diffs staged files against the allowlist	Split the commit or run QA

Verification

• The flag you plan to use still exists: run its Provenance one-liner and get a non-empty hit in the defining file (not only in docs).
• You used the narrowest applicable escape (per-check env var, line directive over file directive).
• The bypass is visible in output (WARN/SKIP line, bypass job summary, or validator notice), not silent.
• If you added a flag: tests cover honored/absent/bad-value, and this catalog has the new row. No manifest bump even when the flag lives in a packaged tree: the manifests carry no version (ADR-092).

Provenance and Maintenance

Audited 2026-07-03 against the working tree. Sources: files and line numbers cited per row above. Line numbers drift; the greps below are the durable re-verification. Run from repo root. If a grep returns nothing, the flag moved or died: update this catalog before relying on it.

Fact	Re-verify one-liner
Git hook jobs, filters, and validators	uv run --frozen lefthook validate
[skip-drift-check] marker + obligations	grep -n "skip-drift-check" .github/workflows/agent-drift-detection.yml
size-exception	grep -n "size-exception" scripts/validation/skill_size.py
orphan-ref directives + 50-line window	grep -n "IGNORE_DIRECTIVE_RE" .claude/skills/orphan-ref-validator/scripts/patterns.py && grep -n "splitlines()\[:50\]" .claude/skills/orphan-ref-validator/scripts/scan.py
investigation allowlist	grep -n "agents/" scripts/modules/investigation_allowlist.py
docs-only verdict	grep -n "SKIPPED: docs-only" .agents/SESSION-PROTOCOL.md
plugin version-field validator	uv run python build/scripts/validate_plugin_version_bump.py --help
no version in any manifest or marketplace entry	uv run python build/scripts/validate_plugin_version_bump.py
GIT_CONFIG_COUNT injection	grep -n "GIT_CONFIG_COUNT" tests/conftest.py
pytest markers	grep -n -A 5 "^markers" pyproject.toml
.env keys	grep -n -e "API_KEY" -e "COMPRESS_TOKENIZER" .env.example
hook registration surfaces	python3 -c "import json; s=json.load(open('.claude/settings.json'))['hooks']; print({k: len(v) for k, v in s.items()})"
removed flags absent from CONTRIBUTING	grep -n -e "SKIP_PREPUSH" -e "SKIP_TESTS" CONTRIBUTING.md (expect no matches)

COMPRESS_TOKENIZER consumer not located; verify before documenting it as live.